車用後視鏡作為汽車的標準配備,隨著台灣汽車銷售趨勢逐漸上升,其需求也隨之增加,廠商對產品品質的把關也越發嚴格。廠內常見的瑕疵類型包括了鏡面刮傷、凸起、凹陷等表面瑕疵,以及缺角、毛邊等輪廓瑕疵。目前業界檢驗方式仍多以人工目視為主,長時間作業下容易因視覺疲勞造成瑕疵分類錯誤率上升。輪廓瑕疵因會造成後視鏡結構的破壞,致使承受應力的能力下降,後視鏡可能會因為嚴重的抖動造成破裂,導致駕駛喪失兩側和後方外部訊息進而引發交通事故。因此,本研究發展一套自動化車用後視鏡輪廓瑕疵檢測系統,以機器視覺取代人員目視檢測。 本研究利用邊點至物件質心的距離對車用後視鏡形狀進行描述,並將此特徵應用於小波轉換搭配低通濾波與指數加權移動平均(EWMA)管制模式檢測瑕疵,該方法透過設定一個合適的EWMA參數,在檢測時不需另外提供一片標準的樣本,從中得到與待測影像比對的資訊,即可運用自身的資訊判斷是否有不規則的輪廓變化,本研究對所提方法分別以車用後視鏡正面與側面影像進行實驗,經由轉換公式計算,能偵測0.2 mm以上瑕疵。實驗結果顯示對於正面影像的正常張數誤判率為7%,瑕疵張數檢出率為86%;側面影像的正常張數誤判率為5%,瑕疵張數檢出率為92%。
In recent years, auto sales gradually increased in Taiwan. Since car mirrors are standard accessories with cars, the demand of car mirrors also increased and manufacturers pay more emphasis on the increase of product quality. Common defects of car mirrors include: scratches, convex, concave causing surface defect type and tailoring, edging negligence of process causing profile defect type. Currently, the inspection tasks are conducted by human inspectors. Since the profile defects will cause structural damages of car mirrors and reduce ability to withstand stress, the degree of harm even more than the surface defects. In additions, the angle diversity of capturing images makes it is not easy to implement automated defect inspection. Therefore, this study develops an automated profile defect detection system of car mirrors to replace visual inspection personnel from car mirror inspection tasks. Two types of images called front images and side images are captured from two views of real car mirror samples. In this study, the distances of edge points to the object centroid are used to describe the shape of a car mirror. To enhance the profile defects on the mirror surface images, the distances of edge points are applied to 1-D wavelet transform with low-pass filtering. The distance deviations of the edge pints before and after the wavelet filtering process can be distinguished by the model of exponential weighted moving average (EWMA) to identify the defect locations. The proposed approach does not require a standard flawless sample in detection process and derive information to compare with testing images. We only use their own information of testing images to determine whether there are any irregular contour changes. The proposed method were carried out to detect the defect size lager than 0.2 mm. Experimental results show that the proposed system achieves 7% false alarm rate and 86% defect detection rate for front image inspection; 5% false alarm rate and 92% defect detection rate for side image inspection.